AI model reconstructs foot-pressure maps from as few as two pressure points

University of Queensland-led researchers reported a multimodal deep-learning model that reconstructed full static plantar-pressure maps from foot geometry and sparse pressure readings. In tests using data from 35 healthy participants, the feature-fusion model recorded normalized RMSE of 0.087 with 16 landmarks and 0.138 with two. The result points to lower-sensor monitoring designs, but the study did not validate a clinical device or test dynamic gait or patient populations.

AI model reconstructs foot-pressure maps from as few as two pressure points University of Queensland-led researchers reported a multimodal deep-learning model that reconstructed full static plantar-pressure maps from foot geometry and sparse pressure readings. In tests using data from 35 healthy participants, the feature-fusion model recorded normalized RMSE of 0.087 with 16 landmarks and 0.138 with two. The result points to lower-sensor monitoring designs, but the study did not validate a clinical device or test dynamic gait or patient populations. A University of Queensland-led research team has developed a deep-learning method that estimates a dense map of pressure across the sole of a foot from its shape and a small set of measured pressure points. The study was published in Sensors on July 1, 2026, and UQ described the work publicly on August 3. The result is an early research finding, not a released medical product. It addresses a practical engineering question: whether an insole could collect fewer measurements while software reconstructs the richer pressure map used in biomechanics and foot-health assessment. How the model works The researchers used static standing data from 35 healthy participants. Their system processed two inputs separately: an image representing plantar geometry and a sparse map of pressure readings at anatomical landmarks. A dual-encoder U-Net then fused those learned features and applied a convolutional block attention module before producing a 64-by-64 pressure map. The team compared feature-level fusion with models that combined the inputs earlier or used only one input type. It also tested configurations with 16, eight, four and two pressure landmarks. Five-fold cross-validation kept all samples from the same participant in one fold, reducing the risk that measurements from one person appeared in both training and validation data. What the reported errors show The feature-fusion model produced the lowest normalized root mean square error across the tested landmark counts. Its reported NRMSE was 0.087 with 16 landmarks and 0.138 with two; lower values indicate closer agreement with the reference pressure maps. The two-landmark result had more error, but it suggests that foot shape and a very small number of localized measurements can still preserve useful information about the broader pressure distribution. That finding could inform lower-cost intelligent insoles or remote assessment tools because fewer physical sensing points may reduce hardware complexity, power needs and cost. It does not show that two sensors are sufficient for diagnosis, treatment decisions or continuous monitoring in real use. The clinical gap remains The experiment covered static standing by healthy participants. It did not evaluate walking, changing loads over time, people with diabetic foot disease or other clinical populations. The authors therefore describe the work as support for future monitoring-system development rather than clinical validation. The paper also discloses industry involvement: two authors were employed by Healthia and one by iOrthotics, and the work was supported by an Australian Cooperative Research Centres Projects grant. For practitioners, the useful takeaway is the architecture and controlled comparison, while product and clinical claims must wait for dynamic, external and patient-population testing. Key Points - 1The model combined plantar geometry with sparse anatomical pressure readings to reconstruct a dense 64-by-64 static pressure map. - 2Its reported normalized RMSE was 0.087 with 16 landmarks and 0.138 with two, with lower error indicating a closer reconstruction. - 3The study used static standing data from 35 healthy participants and did not validate a clinical device, dynamic gait or patient populations. Scoring Rationale A reproducible multimodal-learning result with plausible lower-cost monitoring value, tempered by a small healthy-participant dataset, static-only testing and no clinical-device validation. Sources Primary source and supporting public references used for this report. Practice with real Health & Insurance data 90 SQL & Python problems · 15 industry datasets 250 free problems · No credit card See all Health & Insurance problems /problems/datasets/health